Streaming Fraud Detection
End-to-end real-time credit card fraud detection system. Ingests transaction streams via Apache Kafka, applies a trained Random Forest model through PySpark Structured Streaming, stores predictions in PostgreSQL, and exposes a live Streamlit dashboard for monitoring and analytics.
Project Highlights
Random Forest classifier achieved AUC-ROC: 0.9939, AUC-PR: 0.9597, F1-Score: 0.9042 on live transaction streams.
Built end-to-end pipeline: Kafka ingestion → PySpark feature engineering → PostgreSQL persistence → Streamlit dashboard.
Real-time dashboard with transaction heatmaps, merchant drill-down, and time-series fraud trend visualization.
Fully containerized with Docker Compose for one-command deployment.
Explore More Work
Deep dive into other high-performance solutions.
End-to-end real-time credit card fraud detection system. Ingests transaction streams via Apache Kafka, applies a trained Random Forest model through PySpark Structured Streaming, stores predictions in PostgreSQL, and exposes a live Streamlit dashboard for monitoring and analytics.
Medical information assistant built on Retrieval-Augmented Generation. Accepts PDF and Word medical literature, indexes them with FAISS vector search, and answers clinical questions through a self-reflecting LangChain agent powered by a local Ollama LLM.
Production-ready self-hosted AI automation platform combining n8n, Ollama, Open WebUI, Supabase, and Qdrant — fully containerized with Docker. Enables complex AI workflows locally without cloud dependency, supporting GPU and CPU deployment across Windows, Mac, Linux, and VPS.
AI-powered NPC engine that simulates realistic virtual coworkers for job-skill training simulations. Features dynamic persona registry, internal state tracking (trust, frustration, alignment), conversation memory, and an intelligent pre/post-check supervision layer.
Distributed stock market forecasting pipeline using Apache Spark for feature engineering and three model tiers — Linear Regression, LSTM (PyTorch), and a Hybrid ensemble. Delivers predictions through an interactive Streamlit dashboard with technical indicators and real-time alerts.
Multi-object detection and tracking system for sports video analysis. Uses Faster R-CNN for frame-level detection and Deep SORT for persistent ID assignment across frames — enabling player trajectory extraction and ball tracking for performance analysis.
End-to-end churn analytics solution on 7,032 telecom customer records. Combines EDA, ensemble ML modeling, and a Power BI dashboard to identify at-risk customers, quantify revenue impact, and surface actionable retention strategies.
Multi-label toxic comment classifier using a hybrid CNN + LSTM deep learning architecture. Detects six toxicity categories — toxic, severe toxic, obscene, threat, insult, and identity hate — to automate content moderation at scale.